Talk Title: Optimizing text generation for users: from localizing hallucination detection and attribution towards interaction
Abstract:
As LLMs may hallucinate, users often need to verify the correctness of their generated output. To that end, methods such as attribution, hallucination detection, and more broadly fact checking, were devloped to support users in verifying LLM outputs and detecting their mistakes. However, current such methods are quite hard to use, and often impractical. Attributions typically point at complete documents, or full passages, making it hard to locate the exact evidence spans that support or refute the LLM output. Hallucination detection methods typically point at complete sentences in which some information is suspected as incorrect, but do not pinpoint at the exact mistakes, which usually pertain just to small fractions of a sentence. In this talk I will present a research line whose focus is on localizing the information presented to users, both in hallucination detection — pointing prcisely at the erronous information, and in attribution — where users may verify only specific facts of interest within a sentence and then being presented with the minimal evidence spans that support them. Further, we suggest that such localized hallucination detection and attribution can be integraed under a unifiying framework of comprehensive fine-grained justifications for entailment (NLI) reasoning, at the sub-sentence level. Such framework is proposed not only for facilitating information verification by humans, but also as a detailed feedback mechanism for downstream automated processess that improve LLM factuality, either at training or inference time, as well as for fine-grained evaluation purposes. Finally, I will present briefly a current research line that aims to interactively involve users more deeply in text generation and verification processes, viewed as a human-LLM collaborative writing process.